On the subject of powerful self-improving AI, there does not seem to be enough discussion of real-world limitations or chances for manual override on 1. the AI integrating computational power and more importantly 2. the AI manipulating the outside world with limited info and no dedicated or trustworthy manipulators, or manipulators weaker than Nanotech God. I no longer believe that 1 is a major (or trustable!) limit on FOOM since an AI may be run in rented supercomputers, eat the Internet, etc but 2 seems not to be considered very much. I’ve seen some claims that an AI without too many communication restrictions may be able to anonymously order DNA and stuff and have some idiot mix them and make some kinda self-improving biology up to nanobots, but I haven’t seen anything I could really call a threat assessment.
On the subject of powerful self-improving AI, there does not seem to be enough discussion of real-world limitations or chances for manual override on 1. the AI integrating computational power and more importantly 2. the AI manipulating the outside world with limited info and no dedicated or trustworthy manipulators, or manipulators weaker than Nanotech God. I no longer believe that 1 is a major (or trustable!) limit on FOOM since an AI may be run in rented supercomputers, eat the Internet, etc but 2 seems not to be considered very much. I’ve seen some claims that an AI without too many communication restrictions may be able to anonymously order DNA and stuff and have some idiot mix them and make some kinda self-improving biology up to nanobots, but I haven’t seen anything I could really call a threat assessment.